Meta's Muse Ember targets cheaper, repetitive agent work

Meta AI Research released Muse Ember on November 9, 2026, at $0.10 per million input tokens and $0.50 per million output tokens.

Meta made the model available through Muse Code and the Meta Model API the same day and prices cached input at $0.004 per million tokens.

Meta described Muse Ember as its smallest model and its most efficient option for repeated agent and coding work.

Meta positions the release as a cheaper execution layer beneath its planning and workflow models, with inference cost central to scaling agent deployment.

For Meta, the commercial stakes follow that usage pattern: per-call input and output prices can compound across a multi-step workflow. Meta’s planning-and-execution split makes cost, rather than a single peak intelligence score, its central deployment claim.

What workload is Muse Ember built for?

Meta AI Research says agents spend most of their tokens on small, structured steps: tool calls (commands sent to software), coding subtasks, routing, extraction and long-document parsing.

Ember can retrieve records, call tools in sequence and update systems across applications, Meta says, allowing it to execute a business workflow from start to finish.

Meta says developers can place Ember beneath Muse Wonder’s planning, reducing the number of repetitive calls sent to a more expensive model.

That division follows Meta’s claim that agents spend most of their tokens on repeated, structured execution rather than one large reasoning request.

Meta’s distinction targets the operating layer that businesses repeatedly invoke, not only the planning stage that begins a task.

How does Muse Ember compare with Muse Spark 1.3?

Meta exposes the model as muse-ember-1.0 through the Responses API.

Meta says Ember’s 1M-token context window (the text a model can process in one call) can accommodate a full codebase or document set, while maximum output reaches 512K tokens.

In Meta’s comparison with Muse Spark 1.3, Ember scored 36 on the Artificial Analysis Intelligence Index, versus 37 for Spark 1.3.

Meta lists the cost per Artificial Analysis task at $0.14 for Ember and $2.39 for Spark 1.3.

Meta says Ember is the first model trained using Muse Wonder Full.

Meta’s intended division gives Wonder the hardest planning problems, Spark coding and agentic workflows, and Ember repeatable versions of those tasks at scale.

Meta says Ember matches Claude Haiku 5.5’s best Intelligence Index result at roughly one-third of its cost per task and beats GPT-6 Luna at every reasoning setting.

Meta adds that no model on the full Artificial Analysis leaderboard delivered a higher Intelligence Index score for less per-task cost.

Those cost comparisons are estimates at Meta’s list pricing.

What do the agent and coding benchmarks show?

At maximum reasoning, Meta reports:

Evaluation	Muse Ember	Claude Haiku 5.5	GPT-6 Luna
AutomationBench-AA, business workflows	37.8%	13.9%	20.2%
Agents’ Last Exam, Linux subset	61.6%	57.0%	57.1%
Terminal-Bench 4.0	40.9%	39.2%	16.4%
Coding Agent Index-AA	46	37	31
FrontierCode 2.0-AA	27.2%	34.9%	25.2%
OSWorld 2.1, partial credit	69.5%	68.9%	50.5%

Meta’s table puts Ember ahead on five of six evaluations. Its 37.8% on AutomationBench-AA, 61.6% on Agents’ Last Exam, 40.9% on Terminal-Bench 4.0, 46 on Coding Agent Index-AA and 69.5% on OSWorld 2.1 all lead the comparison models.

FrontierCode 2.0-AA is the exception: Ember’s 27.2% trails Claude Haiku 5.5’s 34.9%, although it exceeds GPT-6 Luna’s 25.2%.

Meta says AutomationBench-AA used strict pass@1 scoring, in which a task must succeed on its first attempt.

Terminal-Bench placed each model in its best published harness, meaning the strongest published execution setup for that model.

Agents’ Last Exam and OSWorld were Meta internal evaluations.

FrontierCode 2.0-AA evaluates real engineering tasks on mergeability, or how well generated code integrates into a codebase; Haiku 5.5 leads it.

Meta’s own FrontierCode result limits the broader coding pitch: efficiency and agent scores do not guarantee superiority on every engineering task.

What can Muse Ember process?

Meta says Ember natively processes text, images, audio and video, making it multimodal (able to handle more than one content format).

An agent can work from screenshots, charts, scanned documents, voice memos, meeting recordings or video without calling a separate model, Meta says.

Meta says those inputs extend Ember’s execution model to desktop computer use and visually detailed documents alongside text-based coding.

What did Meta’s safety testing show?

Meta evaluated Ember under its Superintelligence Scaling Framework for chemical and biological, cybersecurity and loss-of-control risks.

Meta reports that Ember stayed below the framework’s high-risk thresholds across the categories assessed and received the same core safety training and testing.

Meta’s agent-behavior evaluations covered:

• respect for user intent;
• sensitive-information handling;
• resistance to instructions hidden in web pages, documents or tool outputs;
• seeking human input before high-impact actions.

Meta’s checks cover both adherence to user intent and exposure to untrusted instructions in documents, web pages and tool results.

What did early customers report?

Meta published early-testing accounts from Cloudflare and You.com, providing initial outside evidence for Meta’s cost and latency claims.

Cloudflare CTO Dane Knecht said Ember “ran up to 30% faster at the tail” than the comparable model his company tested.

The tail refers to the slowest portion of request latency.

He said Ember also handled investigations spanning multiple repositories well.

You.com inference lead Chak Pothina reported: “95th-percentile latency of 2.2 seconds, with Ember completing every coding quest at the lowest latency among the models tested.”

Here, 95th-percentile latency marks the point below which 95% of request latencies fall.

Pothina also said Ember “produced 36–48% fewer output tokens” during those tests.

Knecht’s and Pothina’s results measure deployment efficiency rather than benchmark scores alone. Cloudflare’s test involved investigations across repositories; You.com’s involved coding quests. Both fit Meta’s stated emphasis on repeated tool use and multi-step work rather than isolated text generation.

Where does Ember fit in Meta’s model stack?

For developers, Meta says, Ember creates a lower-cost execution layer that can sit beneath Muse Wonder’s planning and reduce repetitive calls sent to a more expensive model.

Meta’s cost argument meets a benchmark limit: FrontierCode 2.0-AA gives Claude Haiku 5.5 a 34.9% score against Ember’s 27.2%.

The outside validation remains early: Meta published accounts from Cloudflare and You.com, while Agents’ Last Exam and OSWorld came from Meta internal teams.

Meta’s next test is whether Cloud